In this recorded ControlUp Community Meetup, Chris Twiest explains how the Model Context Protocol—better known as MCP—allows large language models to interact with real systems and real-time data.
MCP is the layer that can turn an LLM from simply answering questions into one capable of investigating environments, correlating telemetry, building reports, creating upgrade plans, and eventually initiating secure actions.
What You’ll Learn:
- Chris begins with a practical introduction to the concepts behind modern AI systems, including:
- Large language models and tokens
- Prompts and system prompts
- Context windows
- Tool calling
- MCP servers
- Agents and copilots
- Retrieval-augmented generation
- Hallucinations and context limitations
He explains why an LLM does not inherently know the current time, understand your environment, or have access to your organization’s data. That information must be supplied through prompts, context, retrieval systems, APIs, and tools.
You will also learn how ControlUp’s AI strategy is evolving from monitoring and Digital Employee Experience toward autonomous endpoint management and automated remediation.
How ControlUp Uses MCP
Chris explains an important distinction between training a model on customer data and allowing a model to securely query that data through tools.
The model is not trained on your ControlUp environment. Instead, the ControlUp MCP Server exposes specific tools that the AI can call when it needs information.
This approach allows the model to retrieve current operational data while ControlUp controls:
- Which tools are available
- Which data sources can be queried
- What permissions the API key provides
- How identifiable information is filtered
- How the model is instructed to interpret the results
Chris also explains why feedback on AI responses matters. Feedback can help the ControlUp team improve system prompts, tool descriptions, data-source selection, and response accuracy without retraining the underlying model on customer data.
Live Demo: Connecting Claude Desktop to ControlUp
The main demonstration walks through connecting Claude Desktop to a local ControlUp MCP Server.
Chris starts with a clean Claude configuration and shows how to:
- Install and use Node.js to run the MCP Server locally
- Generate a ControlUp API key
- Find the ControlUp organization ID
- Add the MCP Server configuration to Claude Desktop
- Restart Claude and confirm that the server is running
- Scan the available ControlUp MCP tools
- Locate the correct application-crash tool
- Retrieve crash data from the previous day
- Visualize the results
- Compare the data against a 14-day baseline
The demonstration shows how quickly raw ControlUp data can be transformed into graphs, trends, summaries, and anomaly analysis through a conversational interface.
Why the System Prompt Matters
Simply giving an AI access to tools does not mean it will always choose the correct tool or interpret the returned data properly.
During the session, Claude initially selects an incorrect data source while attempting to analyze application crashes. It confuses desktop application crashes with errors from a different ControlUp product area.
This is not a problem with access to the data. It is a problem with context.
Chris then supplies a detailed ControlUp system prompt that teaches the model:
- Where different types of ControlUp data are located
- Which tools should be used for particular questions
- How dates and time ranges should be handled
- How application crashes and blue screens are recorded
- How custom queries should be constructed
- How the returned data should be interpreted
With the correct system prompt in place, Claude produces the requested crash analysis more quickly and without becoming distracted by unrelated data sources.
This section provides an excellent practical lesson: MCP gives the AI tools, but the system prompt teaches it how to use them.
Practical MCP Use Cases
- Chris discusses several ways technical teams could use the ControlUp MCP Server:
- Investigate application crashes and performance anomalies
- Compare current telemetry with historical baselines
- Generate custom graphs and HTML reports
- Analyze Windows 10 and Windows 11 readiness
- Build environment-specific upgrade plans
- Evaluate Citrix, Azure Virtual Desktop, VMware, or Windows 365 environments
- Combine ControlUp data with other MCP servers
- Send alerts or results to collaboration platforms
- Create internal AI assistants with custom rules and interfaces
- Build workflows that can eventually perform controlled write actions
For MSPs, multiple MCP Server configurations can be defined using separate API keys and organization IDs for different customers.
Write actions remain governed by the permissions assigned to the API key. The AI cannot perform an action that the associated API credentials are not authorized to perform.
Build Your Own AI Interface
Chris closes with recommendations for anyone who wants greater control over the AI experience.
Instead of relying exclusively on the default Claude Desktop interface, developers can build their own chat interface and control:
- The system prompt
- The model being used
- The available MCP tools
- The supplied context
- The user experience
- The visualization format
- Token consumption
- Data-handling rules
He recommends using Claude Code inside a development environment such as Visual Studio Code or Cursor to build custom AI workflows and interfaces.
One particularly useful technique is to use Claude Desktop to design and refine the detailed prompts that will later be given to Claude Code. This helps provide the coding agent with the context and specifications it needs to produce better results.
Watch the Recording
This session is ideal for ControlUp administrators, EUC professionals, automation engineers, developers, architects, MSPs, and anyone interested in moving beyond basic AI chat toward AI systems that can securely work with live enterprise data.
If you want to understand how MCP works, how ControlUp exposes its data to AI tools, and how you can begin building your own AI-powered operational workflows, watch the complete recording.
Chapters
- 00:08 — Introduction and AI Excitement — Why MCP represents the next evolution of connecting systems, APIs, and AI.
- 05:29 — ControlUp’s AI Vision — The evolution from infrastructure monitoring to DEX, autonomous endpoint management, and automated remediation.
- 08:47 — AI Basics and Core Terms — LLMs, tokens, prompts, system prompts, context windows, retrieval, agents, tool calling, and hallucinations.
- 20:25 — MCP Tools Explained — How MCP gives AI systems access to APIs, tools, and current operational data.
- 25:21 — Training Versus Tools — Why the model is not trained on customer data and how ControlUp securely exposes information through MCP tools.
- 30:06 — How LLMs Use MCP — A technical walkthrough of the interaction between the user, the LLM, the MCP tool, and the underlying API.
- 32:58 — Setting Up Claude — Configuring Claude Desktop, Node.js, the ControlUp API key, organization ID, and local MCP Server.
- 38:23 — First MCP Demo — Retrieving application-crash data, scanning the available tools, and creating a 14-day baseline visualization.
- 44:16 — Fixing the System Prompt — Why the model initially chooses the wrong data source and how a detailed ControlUp system prompt corrects it.
- 48:08 — Better Results with Context — Repeating the analysis with stronger instructions and producing faster, more accurate results.
- 51:39 — Final Tips and Takeaways –Custom chat interfaces, multiple MCP servers, MSP configurations, write actions, self-hosting, and Claude Code.
Please let us know if you have any questions or comments!!
Categories: All Meetups, ControlUp Scripts & Triggers, ControlUp Workflows